Technology-assisted review and predictive coding are machine-learning methods that sort massive volumes of electronically stored information in discovery, training software to recognize which documents are relevant to a case and then applying that judgment across millions of files. Federal courts have endorsed the practice since 2012, and it is now standard in complex commercial disputes, regulatory investigations, and class actions where manual review would be impossibly slow or expensive.
The underlying idea is straightforward. A senior attorney teaches the software what a responsive document looks like, either by coding a sample set up front or by coding documents the system serves up in real time. The algorithm builds a statistical model from those decisions and scores every remaining file. Human review focuses on the documents the model flags, instead of on everything.
How Predictive Coding Actually Works
Two workflows dominate, and the distinction matters because it affects cost, speed, and how the process is defended if challenged.
TAR 1.0: Simple Active Learning
The original approach, Simple Active Learning, begins with a seed set. A subject matter expert, typically a senior attorney with deep knowledge of the case, manually codes roughly 400 to 2,000 documents as relevant or not relevant. That coded sample becomes the training data. The algorithm studies the patterns, assigns a relevance probability score to every file in the collection, and the legal team then cuts the review universe down to the highest-scoring documents after validating accuracy.
The weakness is that the model’s understanding freezes once the seed set is locked. If the case theory shifts or a new custodian introduces unfamiliar vocabulary, responsive files can slip through. Retraining with a new seed set is possible but adds time and cost.
TAR 2.0: Continuous Active Learning
Continuous Active Learning, or CAL, eliminates the formal seed set. The algorithm prioritizes the documents it is least certain about and feeds them to the reviewer in real time. Every coding decision immediately updates the model, so training and review happen at once. CAL tends to find responsive documents faster than TAR 1.0, particularly in collections where relevance is concentrated in a small percentage of files.
It also adapts when new issues surface mid-review. If the expert starts coding for a previously overlooked topic, the algorithm shifts without a formal restart. Most modern TAR platforms now default to CAL.
What Federal Courts Have Said
Three decisions anchor TAR’s legal legitimacy in federal court.
In 2012, Da Silva Moore v. Publicis Groupe became the first federal case to approve predictive coding for document review. The court rejected the notion that manual review is the “gold standard,” stating that “statistics clearly show that computerized searches are at least as accurate, if not more so, than manual review.”1Justia. Da Silva Moore v. Publicis Groupe et al – Document 96
Three years later, Rio Tinto PLC v. Vale S.A. went further, declaring that it was “now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it.” The court emphasized that producing parties are best situated to choose their own review methodology.2Justia. Rio Tinto PLC v. Vale, S.A. et al The approved protocol preserved each party’s right to supplement TAR with keyword searches or other methods.
In 2016, Hyles v. New York City tested the opposite question: can a court force a party to use TAR? The answer was no, but the court expressed a clear preference, noting that “there may come a time when TAR is so widely used that it might be unreasonable for a party to decline to use TAR.” In large-scale disputes, choosing manual review over TAR when proportionality is at issue now invites scrutiny.
When TAR Fits Under Rule 26 Proportionality
Federal Rule of Civil Procedure 26(b)(1) requires that discovery be “proportional to the needs of the case.” Courts weigh the importance of the issues, the amount in controversy, the parties’ relative access to information, their resources, the importance of discovery in resolving the claims, and whether the burden outweighs the likely benefit.3Legal Information Institute. Federal Rules of Civil Procedure Rule 26 – Section: (b) Discovery Scope and Limits
TAR fits this framework because it addresses the burden-versus-benefit question directly. When a collection contains millions of documents, manual review may cost more than the amount in controversy. TAR produces responsive documents at a fraction of that cost, which is the kind of proportional solution the rule contemplates.
Attorneys carry personal exposure here too. Rule 26(g) requires the lawyer’s signature on discovery responses to certify that the response is complete and not unreasonably burdensome. A certification that violates this rule can result in sanctions imposed on the attorney individually, including payment of the opposing party’s reasonable expenses.4Legal Information Institute. Federal Rules of Civil Procedure Rule 26
Building a Defensible Workflow
A TAR process is only as defensible as its documentation. Courts and opposing counsel will scrutinize every decision, from who made the relevance calls to how the training data was assembled.
Choosing the Subject Matter Expert
The subject matter expert is the person whose judgment the algorithm learns to replicate. This should be a senior attorney with firsthand knowledge of the case facts, not a junior associate or contract reviewer. The SME defines what counts as responsive, which means their understanding of the legal claims and defenses directly shapes the algorithm’s output. If the TAR results are challenged, the SME may need to testify about their qualifications, methodology, and the reasoning behind their coding decisions.
Federal Rule of Evidence 702 governs expert testimony and requires the court to find, by a preponderance of the evidence, that the testimony is based on sufficient facts, reliable principles, and a sound application of those principles.5Legal Information Institute. Rule 702 – Testimony by Expert Witnesses An SME who can explain how their experience led to consistent coding decisions will hold up better under cross-examination than one who simply followed the software’s suggestions.
Defining Responsiveness and Assembling Training Data
Before coding begins, the legal team needs a written definition of responsiveness. This might specify date ranges, key custodians, transaction types, or particular topics tied to the claims. Vague definitions produce inconsistent coding, and inconsistent coding produces a model that contradicts itself.
For TAR 1.0 workflows, the SME then codes the seed set. Industry guidelines from the Electronic Discovery Reference Model suggest seed sets typically range from 400 to 2,000 documents, though the ideal size depends on the richness of the collection. The training data should include clear examples of both responsive and non-responsive documents across different file types, custodians, and topics so the algorithm learns broadly rather than fixating on narrow patterns.
For TAR 2.0 workflows, there is no formal seed set. The SME begins reviewing documents the algorithm selects, and each decision feeds back into the model immediately. The advantage is speed. The risk is that the SME must be available consistently throughout the review, since every batch of coding decisions refines the model in real time.
Validation: Proving the Review Worked
An unvalidated TAR process is indefensible. Opposing counsel will ask how you know the algorithm found what it was supposed to find, and “the software said so” is not an answer courts accept. Validation measures two things: whether the algorithm found most of the relevant documents (recall) and whether the documents it flagged were actually relevant (precision).
Recall, Precision, and the F1 Score
Recall measures the proportion of all truly responsive documents in the collection that the algorithm successfully identified. If there are 10,000 responsive documents and the algorithm found 8,000, recall is 80%. Precision is the flip side: of all the documents the algorithm tagged as responsive, what percentage actually were? If the algorithm flagged 12,000 documents but only 8,000 were truly responsive, precision is about 67%.
There is an inherent tension between the two. Pushing recall higher by casting a wider net tends to drag precision down. The F1 score balances them by computing the harmonic mean, giving a single number that reflects overall accuracy. Courts have not mandated a specific recall threshold, but a rate around 75% to 80% is widely regarded as reasonable, and studies have shown that manual human review often falls in the same range or lower.
Elusion Testing
The most common validation method is an elusion test: pulling a random sample from the documents the algorithm classified as non-responsive and having a human check them for missed relevant files. The goal is to confirm that the discard pile contains very few responsive documents. Validation samples are typically drawn at a 95% confidence level with a confidence interval of 2% to 5%, depending on the stakes of the litigation and any agreements with opposing counsel.
If the elusion test reveals an unacceptable number of missed documents, the team feeds additional examples into the model and retrains. For TAR 2.0, this happens organically as coding continues. For TAR 1.0, it may require assembling a new seed set and restarting the training cycle, which is one reason CAL has largely overtaken the original approach.
Protecting Privilege With a Rule 502(d) Order
One of the biggest risks in any large-scale production is accidentally handing over privileged material. When millions of files move through an automated pipeline, even a well-trained algorithm can miss an attorney-client email buried in a forwarded chain. Federal Rule of Evidence 502 was designed to reduce this risk.
Under Rule 502(b), an inadvertent disclosure during a federal proceeding does not waive the privilege if three conditions are met: the disclosure was genuinely inadvertent, the producing party took reasonable steps to prevent it, and the party acted promptly to fix the error once discovered. The advisory committee note explicitly contemplates TAR, stating that “a party that uses advanced analytical software applications and linguistic tools in screening for privilege and work product may be found to have taken ‘reasonable steps’ to prevent inadvertent disclosure.”6Legal Information Institute. Rule 502 – Attorney-Client Privilege and Work Product; Limitations on Waiver
The strongest protection comes from a court order under Rule 502(d), which provides that any disclosure connected with the litigation does not constitute a waiver, whether the disclosure was inadvertent or intentional. Unlike a private agreement, a 502(d) order is binding in all other federal and state proceedings, so a privilege slip in one case cannot be used against the party in a later lawsuit.6Legal Information Institute. Rule 502 – Attorney-Client Privilege and Work Product; Limitations on Waiver Experienced e-discovery practitioners treat a 502(d) order as non-negotiable in any TAR-assisted production, and most courts grant them routinely.
Even with the order in place, the producing party still runs a privilege screen before production. Clawback provisions establish what happens when something slips through: the producing party sends written notice identifying the document, the receiving party must return or destroy all copies, and any dispute gets resolved by the court after reviewing the document privately. The document cannot be used in the litigation while the dispute is pending.
What TAR Costs
The economics are TAR’s strongest selling point. Contract attorneys performing manual review typically charge between $25 and $65 per hour, and even at those rates a collection of five million documents can generate review costs that exceed the amount in controversy. TAR doesn’t eliminate human review, but it dramatically reduces the volume of documents that need human eyes.
Processing costs for converting raw data into a reviewable format generally run between $25 and $100 per gigabyte, depending on file-type complexity and platform. Overall e-discovery spend using TAR is routinely 50% to 80% less than a comparable manual review would cost for matters involving several million documents.
The savings are most dramatic in cases with low richness, where only a small fraction of collected documents are actually responsive. In a collection where 2% of files are relevant, TAR can surface that 2% without requiring humans to slog through the other 98%. Manual review of the same collection would mean paying reviewers to look at every document, most of which are irrelevant.
The calculation gets trickier in smaller matters. If the collection is only 50,000 documents, the overhead of setting up a TAR workflow, selecting an SME, running validation, and preparing a protocol may not save money compared to a focused keyword search followed by manual review. TAR’s sweet spot is collections too large for humans to review efficiently but important enough to justify the upfront investment in training and validation.
The Attorney’s Duty of Technological Competence
Lawyers who use TAR carry ethical duties beyond running the software correctly. Comment 8 to ABA Model Rule 1.1 requires lawyers to “keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology.”7American Bar Association. Rule 1.1 Competence – Comment At least 40 states have adopted this duty of technological competence, so ignorance of tools like TAR is itself an ethical problem for attorneys handling large-scale discovery.
The obligation does not mean every lawyer needs to become a data scientist. It means that an attorney responsible for a document production must understand enough about the chosen technology to supervise it meaningfully. Handing a collection and a discovery request to a vendor and asking them to handle it crosses the line into delegating legal judgment to a non-lawyer. The attorney must direct the review, define the criteria for responsiveness, and verify the results. The vendor handles the technical infrastructure; the lawyer makes the legal calls.
When outside vendors are involved, the supervising attorney also needs to confirm that the vendor’s processes do not expose client data to unnecessary risk. Entering confidential case information into publicly accessible AI tools may violate the duty of confidentiality.
Where Generative AI Fits In
Large language models are beginning to reshape document review. Unlike traditional TAR, which requires either a manually coded seed set or iterative reviewer feedback, generative AI models can classify documents using natural-language instructions alone. A reviewer can describe what responsive documents look like in a few sentences, and the model applies that description across the collection without coded training examples. Research has shown this zero-shot approach can produce accurate annotations without the manual setup TAR 1.0 requires.8Frontiers in Artificial Intelligence. The Unreasonable Effectiveness of Large Language Models in Zero-Shot Semantic Annotation of Legal Texts
Courts are paying attention. A growing number of federal judges have issued standing orders requiring attorneys to disclose whether they used generative AI tools in preparing filings or managing discovery. Some orders require identification of the specific tool and how its output was incorporated into the work product. Others mandate that all AI-generated content be verified by a licensed attorney using non-AI sources before submission.9Kathrine R. Everett Law Library. Judicial Guidance on the Use of GenAI in Court The regulatory picture is unsettled. The Fifth Circuit declined to adopt a proposed rule requiring AI certification, while at least one state supreme court has discussed a potential outright ban on AI in legal proceedings.
The practical concern is confidentiality. Feeding privileged client documents into a cloud-based language model raises serious questions about whether the data remains protected. Some courts have explicitly warned that entering confidential information into public AI tools may violate professional obligations.9Kathrine R. Everett Law Library. Judicial Guidance on the Use of GenAI in Court Most legal teams using generative AI for document review are doing so through private, enterprise-grade deployments that keep client data within controlled environments. Checking the local rules of the specific court before review begins is essential, because disclosure requirements are evolving quickly.